Pre-reaching infants expect causal agents to reach efficiently
Shari Liu
July 31 2018
Tables and Figures
Figure 1. Still frames from videos shown to participants in Experiments 1-5, including stimuli from habituation (A) and test (B). In each video, a person reached for and caused a change in an object (H1-H3, T1-T2), or picked up the object (H4-H5, T3-T4), over a barrier (H1- H2, H4-H5) or over empty space (H2, T1-T4). The person either acted on the object by contacting it (H1-H2, H4-H5, T1, T3-T4) or produced the same effect from a distance of 50 pixels, after a 0.5s delay (H3, T2), and either performed these actions while wearing a mitten (H1-H4, T1-T3) or with a bare hand (H5, T4) During test (B), the person either reached directly for the object on a novel but efficient trajectory (left panels), or in a curvilinear fashion on the familiar but inefficient trajectory (right panels).
Figure 2. Looking time in seconds towards the efficient versus inefficient reach (bottom), and proportion looking towards the inefficient reach (top) at test across Experiments 1-5 (N=152). Labels above each panel list the experiment name (Exp. 1-5), whether actions during habituation were constrained or unconstrained by a barrier, goal (state.change or pick.up), whether these actions involved contact with the object, whether the actor wore a mitten, and video displays listed in Figure 1. Error bars around means indicate within-subjects 95% confidence intervals (bottom) and bootstrapped 95% confidence intervals (top). Individual points (top) or pairs of points (bottom) indicate data from a single participant. Horizontal bars within boxes indicate medians, and boxes indicate the middle 2 quartiles of data. Violin plots (top) indicate distribution of data, area scaled proportionally to the number of observations.
Results (Main Text)
Infants’ analysis of causal vs non-causal actions
Experiment 1
In Experiment 1 (N=40; 20 per condition), 3-month-old infants (M=107.675 days, range=91-122, 23 female) were randomly assigned to two conditions. Infants in the constrained condition viewed repeated video clips of a person wearing mittens who reached over a barrier that prevented direct access to an object and appeared to cause the object to light up by touching it (Figure 1A, H1). Infants in the unconstrained condition viewed the same reaches with the barrier behind the goal object, out of the actor’s way (H2): an action that appears to adults and older infants to be inefficient. In both conditions, the object illuminated on contact with the person’s hand. Once infants were habituated to (i.e. lost interest in) these repeating events, they were tested with alternating events in which the person either reached on a previously seen, curved trajectory (a familiar but now inefficient action) or on a direct path to the object (a novel but efficient action) (Figure 1B, T1). Across all experiments, we calculated the average looking time towards the efficient versus inefficient reach over 3 pairs of test events. Infant looking times are often log-normally distributed, including in this dataset (see SI) and thus were log-transformed (main results) or transformed to proportions (meta-analysis, see SI) prior to analysis. See Methods for details.
A mixed effects model, including a fixed interaction between habituation and test event and a random intercept for participant identity, revealed that infants responded differently to the test events across the two conditions, [0.273,0.732], ß=0.781, B(SE)=0.502(0.114), p<.001, two-tailed, 2 participants excluded on the basis of Cook’s Distance. Contrasts from this model revealed that when the actor’s reaches were initially constrained by a barrier (H1), infants looked longer at the inefficient action (M=15.448s, SE=0.658) than the efficient action (M=12.368s, SE=0.658) at test, [0.396,1.139], ß=1.194, B(SE)=0.768(0.185), p<.001, two-tailed. Critically, this looking preference cannot be attributed to low-level responses to the curvilinear reach, because infants in the unconstrained condition (H2) showed a preference for the efficient (M=10.104s, SE=0.738) over the inefficient (M=8.788s, SE=0.738) action, [-0.343,-0.017], ß=-0.28, B(SE)=-0.18(0.081), p=0.032, two-tailed. See Figure 2. See Figure 2. Experiment 1 therefore provides evidence that infants expected the actor to reach efficiently.
Experiment 2
In Experiment 2 (N=20), we asked whether this expectation depended on infants’ construal of the actor as a causal agent, who aimed to change the state of the object on contact.We manipulated the spatiotemporal continuity of the reaching action: a variable that modulates causal perception in older infants and adults and in computational studies of learning about objects and actions. Infants (M=106.7 days, range=93-121, 12 female) were habituated to videos identical to H1, except that the person’s hand stopped 50 pixels above the object (approximately 2 cm on the presentation screen), and the object illuminated after a 0.5 second delay. At test, the person reached efficiently or inefficiently in the absence of the barrier, with the same spatiotemporal gap as in habituation trials (T2). I nfants looked equally between the inefficient (M=15.306s, SE=1.123) and efficient (M=16.38s, SE=1.123) reach, [-0.191,0.301], ß=0.096, B(SE)=0.055(0.119), p=0.649, two-tailed, mixed effects model with trial type as fixed effect and participants as a random intercept. Their looking patterns differed from those of infants shown causally effective reaches in the corresponding condition of Experiment 1, [-0.623,-0.003], ß=-0.547, B(SE)=-0.313(0.154), p=0.049, two-tailed, mixed effects model with fixed interaction between causality and trial type and participants as a random intercept. This finding suggests that infants view action cost and causal efficacy as linked properties of agents’ actions.
Experiment 3
To evaluate this suggestion further, Experiment 3 (N=52; 26 per condition), pre-registered at https://osf.io/f2hvd/, attempted to replicate the findings from Experiments 1 and 2 by randomly assigning infants (M=107.154 days, range=92-121, 21 female) to two conditions manipulating causal information alone: a design that allowed us to compare infants’ expectations about efficiency to causal vs non-causal actions, as assessed by experimenters who were blind to condition as well as test events. Its findings replicated those of the previous two studies: Infants viewed the agent’s actions as efficiently directed to the goal object only when her hand contacted the object at the time of the object’s state change, [-0.815,-0.184], ß=-0.729, B(SE)=-0.5(0.158), p=0.003, two-tailed, mixed effects model with fixed interaction of causality and test event and random intercept for participants. As in Experiment 1, infants looked longer at the inefficient reach (M=12.166s, SE=0.588) than the efficient reach (M=7.791s, SE=0.588) when the person’s actions were causally effective, [0.211,0.66], ß=0.635, B(SE)=0.436(0.112), p<.001, one-tailed, and as in Experiment 2, infants looked equally to the inefficient (M=11.395s, SE=0.818) and efficient (M=12.888s, SE=0.818) reaches when the spatiotemporal gap between her hand and the object’s illumination suggested that she did not cause this event, [-0.289,0.16], ß=-0.094, B(SE)=-0.064(0.112), p=0.284, one-tailed. Together, Experiments 1-3 show that pre-reaching infants apply the principle of efficiency to reaching actions that they themselves have never experienced, when the actions appear to cause changes in the target object. Together, Experiments 1-3 show that pre-reaching infants apply the principle of efficiency6 to reaching actions that they themselves have never experienced, provided that the actions occur on contact with the target object. Although 3-month-old infants have limited experience acting as causal agents themselves and likely no experience illuminating objects by touching them, they apply a causal analysis to this action when it is performed by others.
Infants’ analysis of entrainment actions
Experiments 4 and 5
In Experiment 4 (N=20), infants (N=20), infants (M=107.9 days, range=92-122, 11 female) were habituated to and tested on events in which a person reached for and picked up an object identical to those from Experiments 1-3 while wearing the same mitten (H4, T3), as in prior research. In Experiment 5 (N=20), infants (M=107.95 days, range=93-120, 12 female) saw almost identical videos to those from Experiment 4 except that the person reached with a bare hand (H5, T4). Infants looked longer at the inefficient (M=9.715s, SE=0.474) than the efficient (M=8.036s, SE=0.474) reach of the bare hand, [0.008,0.331], ß=0.296, B(SE)=0.17(0.08), p=0.02, one-tailed, but did not distinguish the inefficient (M=18.029s, SE=0.643) from the efficient (M=16.844s, SE=0.643) action of the mittened hand [-0.083,0.232], ß=0.13, B(SE)=0.074(0.078), p=0.172, one-tailed. However, infants’ looking patterns in Experiments 4 and 5 did not differ from each other, [-0.128,0.319], ß=0.167, B(SE)=0.095(0.111), p=0.396, two-tailed, mixed effects model with fixed interaction between mitten and test trial, random intercept for participants, excluding 3 influential participants on the basis of Cook’s Distance. Thus, Experiment 5 provides evidence that infants expect barehanded reaching and lifting of an object to be efficient, but the differences in infants’ responses to gloved and ungloved hands that pick up an object are not clear.
Comparing state change and entrainment actions
To explore this and other effects further, in relation to the effects obtained in past research32, we performed an analysis over the ten experiments in these two papers (total N=264 pre-reaching infants, 12 conditions). Our analytic approach allows us to assess the independent effects of 5 manipulations: the type of or absence of motor training, the presence or absence of barrier preventing a direct reach for the object during habituation, the nature of the goal (to change the state of an object or pick it up), the presence or absence of action on contact, and the presence of absence of mittens on the actor. The analysis also allows us to control for the participant variables age and sex, and model the nested structure of the data (e.g. looks clustered within experiments and within papers). For ease of interpretation, we used average proportion looking to the inefficient action in this analysis. Infants expected efficient reaching in all conditions in which the agent previously had reached efficiently over a barrier (relative to no barrier) and contacted the object (relative to no contact), but they held stronger expectations for efficient reaching when the reach resulted in a simple contact and state change of an object than when it lifted and entrained the object, [0.02,0.053], ß=0.397, B(SE)=0.035(0.01), p=0.02, two-tailed. For full meta-analytic results, see SI. Knowledge of the causal mechanisms underlying complex actions therefore becomes more robust when infants perform these actions, but sensitivity to abstract properties of actions predates this learning.
Reliability (Methods Section)
To assess reliability, 50% of test trials from participants across Experiments 1-5 (132 participants, 456 trials) were randomly selected and coded by additional researchers who were unaware of experimental condition, and test trial order. The intraclass correlation coefficient (ICC) between the original data, and this newly coded data, was 0.968 [0.955, 0.978], 0.963 [0.938, 0.977], 0.936 [0.911, 0.954], 0.969 [0.946, 0.982], 0.969 [0.943, 0.982], for Experiments 1 through 5, respectively.
Supplemental Results
Comparing Experiments 4-5 with Skerry et al. (2013)
We compared the results of Experiment 4 and 5 against those from Skerry et al’s Experiment 3, wherein infants received no mittens training and viewed a person reaching with a mittened hand. The results of Experiment 5 (no mitten) differed from those of the earlier experiment (mitten), [0.047,0.547], ß=0.539, B(SE)=0.297(0.124), p=0.022, two-tailed, mixed effects model with fixed interaction between experiment and test event and random intercept for participants, one influential participant excluded on the basis of Cook’s Distance. In addition, the results from Experiment 4 (mitten) marginally differed from those in of Skerry et al.32 (mitten), [-0.021,0.47], ß=0.43, B(SE)=0.224(0.122), p=0.074, two-tailed, mixed effects model with fixed interaction between experiment and test events and random intercept for participants, 2 influential participants excluded on the basis of Cook’s Distance.
Meta-analytic results
To assess the unique effects of our experimental manipulations in Experiments 1-5 and in Skerry et al.32 , we performed an analysis over these two papers (total N=264, 12 conditions). We assessed the independent effects of 5 manipulations (see Figure S1) while controlling for participant variables age and sex and modeling the nested structure of the data. For ease of interpretation, we used average proportion looking to the inefficient action in this analysis, following Skerry et al.3
This analysis confirmed that first-person action experience is not the only way to enhance infants’ appreciation of the causal and intentional aspects of action. It also confirmed the findings from the individual experiments reported in the main text and from SCS: Infants’ expectations were stronger when the observed action was spatiotemporally continuous its effect (i.e., appeared to be causal), [0.027,0.06], ß=0.501, B(SE)=0.044(0.009), p<.001, two-tailed, when infants received effective motor training (sticky mittens), relative to no training [0.027,0.069], ß=0.558, B(SE)=0.049(0.011), p=0.003, two-tailed, when the observed agent’s actions were constrained by a barrier and were efficiently adapted to that barrier, relative to the same actions that were unconstrained by a barrier, [0.021,0.051], ß=0.407, B(SE)=0.036(0.008), p=0.001, two-tailed, and when the agent pursued a state change goal, relative to a pickup goal, [0.02,0.053], ß=0.397, B(SE)=0.035(0.01), p=0.02, two-tailed. We also found that infants’ expectations were marginally negatively affected when they received ineffective motor training (non-sticky mittens), relative to no training, [-0.06,-0.005], ß=-0.354, B(SE)=-0.031(0.015), p=0.068, two-tailed, and were unaffected when the actor wore a mitten, relative to no mitten [-0.045,0], ß=-0.232, B(SE)=-0.021(0.012), p=0.14, two-tailed, as reported in the main text. These findings provide further evidence that action experience alters action interpretation, for good or for ill, but so does causal information and information about efficiency.
Figure S1. Looking time in seconds towards the efficient versus inefficient reach (bottom), and proportion looking towards the inefficient reach (top) at test across Experiments 1-5 (n=152) and across Experiments 1-5 in Skerry et al. (SCS)32 (n=112). Labels above each panel list the experiment name (Exp. 1-5, SCS Exp. 1-5), type of motor training (none, ineffective non-sticky mittens, or effective sticky mittens), whether actions during habituation were constrained or unconstrained by a barrier, goal (state.change or pick.up), whether actions resulted in contact with the object, whether the actor wore a mitten, and video displays listed in Figure 1. Error bars around means indicate within-subjects 95% confidence intervals (bottom) and bootstrapped 95% confidence intervals (top). Individual points (top) or pairs of points (bottom) indicate data from a single participant. Horizontal bars within boxes indicate medians, and boxes indicate the middle 2 quartiles of data. Violin plots (top) indicate distribution of data, area scaled proportionally to the number of observations.
Figure S2. Effect plots for model investigating predictors of sensitivity to action efficiency across Experiments 1-5 and Skerry et al. (2013) 32 (total N=264, 247 included in final analysis, 17 excluded on the basis of Cook’s Distance). Each point shows estimates of effects at each level of all categorical predictors: Type of motor training (none, ineffective non-sticky mittens, or effective sticky mittens), the goal of the actor (state change vs pick up), action during habituation (constrained or unconstrained by a barrier), whether actions resulted in contact with the object (yes or no), whether the actor wore a mitten (yes or no). Error bars indicate 95% confidence intervals. See Table S1 for full results.
Table S1. Regression table for model investigating predictors of sensitivity to action efficiency across Experiment 1-5 and all experiments from Skerry et al. (total N=264, 247 included in final analysis, 17 excluded on the basis of Cook’s Distance). Dependent measure is proportion looking towards the inefficient reach, averaged across 3 test trials during test. Categorical predictors were coded using sum contrasts, and fixed effects from the model should therefore be interpreted with respect to the grand mean (with respect to 0). Model formula: prop.ineff.all ~ training + goal + hab + causal + mitten + (1|experiment) + (1|ageday) + (1|sex) + (1|paper).
| Standardized Estimate (ß) | Estimate (B) | Standard Error (SE) | df | t | p | 95% CI (Lower) | 95% CI (Upper) | |
|---|---|---|---|---|---|---|---|---|
| (Intercept) | -0.340 | 0.488 | 0.019 | 2.19 | 25.28 | 0.001 | 0.457 | 0.523 |
| effective training | 0.558 | 0.049 | 0.011 | 7.32 | 4.31 | 0.003 | 0.027 | 0.069 |
| ineffective training | -0.354 | -0.031 | 0.015 | 8.70 | -2.08 | 0.068 | -0.060 | -0.005 |
| state change goal | 0.397 | 0.035 | 0.010 | 4.24 | 3.66 | 0.020 | 0.020 | 0.053 |
| constrained habituation | 0.407 | 0.036 | 0.008 | 9.61 | 4.54 | 0.001 | 0.021 | 0.051 |
| causally effective | 0.501 | 0.044 | 0.009 | 20.54 | 5.08 | 0.000 | 0.027 | 0.060 |
| mitten | -0.232 | -0.021 | 0.012 | 7.39 | -1.65 | 0.140 | -0.045 | 0.000 |
This analysis confirmed that first-person action experience is not the only way to enhance infants’ appreciation of the causal and intentional aspects of action. It also confirmed the findings from the individual experiments reported in the main text and from Skerry et al. (2013): Infants’ expectations were stronger when the observed action was spatiotemporally continuous with its effect (i.e., appeared to be causal),[0.027,0.06], ß=0.501, B(SE)=0.044(0.009), p<.001, two-tailed, when infants received effective motor training (sticky mittens), relative to no training, [0.027,0.069], ß=0.558, B(SE)=0.049(0.011), p=0.003, two-tailed, when the observed agent’s actions were constrained by a barrier and were efficiently adapted to that barrier, relative to the same actions that were unconstrained by a barrier, [0.021,0.051], ß=0.407, B(SE)=0.036(0.008), p=0.001, two-tailed, and when the agent pursued a state change goal, relative to a pickup goal, [0.02,0.053], ß=0.397, B(SE)=0.035(0.01), p=0.02, two-tailed. We also found that infants’ expectations were marginally negatively affected when they received ineffective motor training (non-sticky mittens), relative to no training, [-0.06,-0.005], ß=-0.354, B(SE)=-0.031(0.015), p=0.068, two-tailed, and were unaffected when the actor wore a mitten, relative to no mitten [-0.045,0], ß=-0.232, B(SE)=-0.021(0.012), p=0.14, two-tailed, as reported in the main text.
Exclusion info
Table S2. Tally of infants who participated in Experiments 1-5 but were excluded in our final sample. These exclusion criteria were set prior to the start of data collection for each experiment, but vary slightly across experiments (e.g. we relaxed our definition of inattentiveness from excluding all data from a participant if they missed a test trial in Experiment 1, to excluding data from just that trial in Experiments 2-5).
| Experiment | Fussiness | Inattentiveness | Caregiver Interference | Experimenter/Coding Error | Technical Failure | Total |
|---|---|---|---|---|---|---|
| Exp.1 | 9 | 5 | 1 | 12 | 3 | 30 |
| Exp.2 | 0 | 0 | 0 | 2 | 0 | 2 |
| Exp.3 | 6 | 0 | 0 | 2 | 0 | 8 |
| Exp.4 | 7 | 0 | 0 | 2 | 0 | 7 |
| Exp.5 | 6 | 0 | 0 | 1 | 2 | 9 |
| Total | 28 | 5 | 1 | 19 | 5 | 50 |
Distribution of Looks
Figure S3. Density plot of looking times during test across Experiments 1-5, and Experiments 1-5 from Skerry et al. (2013) (N=264). Maximum-likelihood fitting revealed that the lognormal distribution (log likelihood=-1720.509) provides a better fit to these data than the normal distribution (log likelihood=-1842.196).
Discrepancy between codings of data
Figure S4. Discrepancy between original and new coding of looking times during test trials in seconds. For tests of reliability, see Methods.
Attention during habituation across Exp 1-5
Figure S5. Total looking time in seconds during habituation across Experiment 1-5. Error bars around means indicate bootstrapped 95% confidence intervals. Individual points indicate data from a single participant. Horizontal bars within boxes indicate medians, and boxes indicate the middle 2 quartiles of data. Violin plots in indicate distribution of data, area scaled proportionally to the number of observations.
Figure S6. Looking time in seconds during each habituation trial across Experiments 1-5. Curves with 95% confidence interval ribbons indicate smoothed conditional means, generated using the loess method. Connected points indicate data from a single participant. Labels above each panel list the experiment name (Exp. 1-5), whether actions during habituation were constrained or unconstrained by a barrier, goal (state.change or pick.up), whether actions resulted in contact with the object, whether the actor wore a mitten, and video displays listed in Figure 1.
Tables S4-S8 Regression table for mixed effects model analyzing the effect of age, sex, order of test events, habituation condition, goal, mitten, and causal information on habituation, controlling for variations across Experiments 1-5. Model formula: total_hab ~ ageday + sex + first.test + hab + goal + mitten + causal + (1|experiment)
| Standardized Estimate (ß) | Estimate (B) | Standard Error (SE) | df | t | p | 95% CI (Lower) | 95% CI (Upper) | |
|---|---|---|---|---|---|---|---|---|
| (Intercept) | -0.208 | 343.171 | 76.54 | 151.82 | 4.483 | 0.000 | 192.192 | 494.168 |
| Age in Days | -0.233 | -2.058 | 0.68 | 147.57 | -3.026 | 0.003 | -3.400 | -0.714 |
| Sex | 0.066 | 5.203 | 6.11 | 148.69 | 0.852 | 0.396 | -6.916 | 17.274 |
| First Test Event | -0.006 | -0.439 | 6.00 | 146.69 | -0.073 | 0.942 | -12.270 | 11.393 |
| Habituation | 0.222 | 17.590 | 11.03 | 131.60 | 1.595 | 0.113 | -6.441 | 41.220 |
| Goal | 0.007 | 0.589 | 16.02 | 6.18 | 0.037 | 0.972 | -0.777 | 50.348 |
| Mitten | 0.126 | 9.996 | 19.02 | 5.73 | 0.525 | 0.619 | 8.336 | 50.922 |
| Causal | -0.055 | -4.379 | 9.08 | 75.38 | -0.482 | 0.631 | -5.024 | 3.212 |
To ask whether infants’ total attention during habituation was affected by experimental manipulations across Experiment 1-5 (action constrained vs unconstrained by a barrier, state change vs pickup goal, mitten vs no mitten on actor, and action with vs without contact with the object), and varied by gender and age, we fit a mixed effects model on these fixed effects and experiment (Exp.1-5) as a random intercept. We found that the only robust predictor of attention during habituation was age, [-3.4,-0.714], ß=-0.233, B(SE)=-2.058(0.68), p=0.003, two-tailed, such that older infants looked for a shorter time overall than younger infants.